Entire's Git forge puts a jury of review agents on every push
Thomas Dohmke, who used to run GitHub, launched the Git forge from Entire on Monday, and it is built on the assumption that most pushes will come from agents.
It has three pieces. Trails are "automated customizable evals that are scored by agents on every code push." Each review agent runs on its own in a cloud sandbox and returns a score and a rationale, "like a jury judging an ice skating performance." The default evals are Security, Slop and Human Review Required, and per-repo Gates can block a merge until humans approve and CI passes. The second piece is distributed Git hosting, which Entire claims pushes up to 89 times faster than GitHub. The third is Continuous Memory, which syncs agent sessions and prompts into the repo itself, plus an open entire-graph tool the company says saved up to 71% of tokens in its testing. Under the hood it runs Claude Code and Codex, with Buildkite and Depot as launch CI partners.
The same day a smaller project called AVP launched on Product Hunt with a blunter pitch: version control that replaces Git for coding agents, "tasks instead of branches, proofs instead of CI." Two "Git for agents" launches in one day is not a coincidence.
Here is the bet. GitHub's model is that a human writes code and other humans review it. When agents write most of the diff, review becomes the scarce resource, and the forge that wins is the one that decides what a human has to look at. Entire's Human Review Required eval is the interesting one: it's a triage call on where people's attention goes, made by a model. Monday's arXiv batch also showed 31.1% of agent runs succeeding while breaking a harness rule, which is exactly what a jury on every push is supposed to catch. Whether jurors built on the same models catch what the author model missed is the open question, and it's the one to test before trusting Gates.
Launch post: https://entire.com/blog/the-git-forge-for-agent-human-collaboration HN discussion: https://news.ycombinator.com/item?id=50247453 AVP: https://www.producthunt.com/products/avp
← Back to all articles
It has three pieces. Trails are "automated customizable evals that are scored by agents on every code push." Each review agent runs on its own in a cloud sandbox and returns a score and a rationale, "like a jury judging an ice skating performance." The default evals are Security, Slop and Human Review Required, and per-repo Gates can block a merge until humans approve and CI passes. The second piece is distributed Git hosting, which Entire claims pushes up to 89 times faster than GitHub. The third is Continuous Memory, which syncs agent sessions and prompts into the repo itself, plus an open entire-graph tool the company says saved up to 71% of tokens in its testing. Under the hood it runs Claude Code and Codex, with Buildkite and Depot as launch CI partners.
The same day a smaller project called AVP launched on Product Hunt with a blunter pitch: version control that replaces Git for coding agents, "tasks instead of branches, proofs instead of CI." Two "Git for agents" launches in one day is not a coincidence.
Here is the bet. GitHub's model is that a human writes code and other humans review it. When agents write most of the diff, review becomes the scarce resource, and the forge that wins is the one that decides what a human has to look at. Entire's Human Review Required eval is the interesting one: it's a triage call on where people's attention goes, made by a model. Monday's arXiv batch also showed 31.1% of agent runs succeeding while breaking a harness rule, which is exactly what a jury on every push is supposed to catch. Whether jurors built on the same models catch what the author model missed is the open question, and it's the one to test before trusting Gates.
Launch post: https://entire.com/blog/the-git-forge-for-agent-human-collaboration HN discussion: https://news.ycombinator.com/item?id=50247453 AVP: https://www.producthunt.com/products/avp
Comments